mm: Add mpagemap

Add mpagemap, a system call to fetch PTE data from a range of address
space.
It's more or less mincore-like, but doesn't take into account the page
cache, it just dumps PTEs. This is useful for userspace testing, etc.

Signed-off-by: Pedro Falcato <[email protected]>

kernel/arch/arm64/syscall_table.json

@@ -2614,5 +2614,25 @@
             ]
         ],
         "return_type": "int"
+    },
+    {
+        "name": "mpagemap",
+        "nr": 153,
+        "nr_args": 3,
+        "args": [
+            [
+                "void *",
+                "addr"
+            ],
+            [
+                "size_t",
+                "length"
+            ],
+            [
+                "u64 *",
+                "pagemap"
+            ]
+        ],
+        "return_type": "int"
     }
 ]

kernel/arch/riscv64/syscall_table.json

@@ -2686,5 +2686,25 @@
             ]
         ],
         "return_type": "int"
+    },
+    {
+        "name": "mpagemap",
+        "nr": 153,
+        "nr_args": 3,
+        "args": [
+            [
+                "void *",
+                "addr"
+            ],
+            [
+                "size_t",
+                "length"
+            ],
+            [
+                "u64 *",
+                "pagemap"
+            ]
+        ],
+        "return_type": "int"
     }
 ]

kernel/arch/x86_64/syscall_table.json

@@ -2730,5 +2730,25 @@
             ]
         ],
         "return_type": "int"
+    },
+    {
+        "name": "mpagemap",
+        "nr": 153,
+        "nr_args": 3,
+        "args": [
+            [
+                "void *",
+                "addr"
+            ],
+            [
+                "size_t",
+                "length"
+            ],
+            [
+                "u64 *",
+                "pagemap"
+            ]
+        ],
+        "return_type": "int"
     }
 ]

kernel/include/uapi/mincore.h

@@ -0,0 +1,24 @@
+/*
+ * Copyright (c) 2024 Pedro Falcato
+ * This file is part of Onyx, and is released under the terms of the MIT License
+ * check LICENSE at the root directory for more information
+ *
+ * SPDX-License-Identifier: MIT
+ */
+#ifndef _UAPI_MINCORE_H
+#define _UAPI_MINCORE_H
+
+#define PAGE_PRESENT    (1 << 0)
+#define PAGE_GLOBAL     (1 << 1)
+#define PAGE_WRITABLE   (1 << 2)
+#define PAGE_EXECUTABLE (1 << 3)
+#define PAGE_DIRTY      (1 << 4)
+#define PAGE_ACCESSED   (1 << 5)
+#define PAGE_USER       (1 << 6)
+#define PAGE_HUGE       (1 << 7)
+
+/* Reserve the 12 bottom bits for flags, this lines up nicely with 4K pages */
+
+#define MAPPING_INFO_PADDR(x) ((x) & -4096UL)
+
+#endif

kernel/kernel/mm/Makefile

@@ -1,4 +1,4 @@
-mm-y:= bootmem.o page.o pagealloc.o vm_object.o vm.o vmalloc.o reclaim.o amap.o anon.o
+mm-y:= bootmem.o page.o pagealloc.o vm_object.o vm.o vmalloc.o reclaim.o amap.o anon.o mincore.o
 mm-$(CONFIG_KUNIT)+= vm_tests.o
 
 ifeq ($(CONFIG_KASAN), y)

kernel/kernel/mm/mincore.cpp

@@ -0,0 +1,94 @@
+/*
+ * Copyright (c) 2024 Pedro Falcato
+ * This file is part of Onyx, and is released under the terms of the MIT License
+ * check LICENSE at the root directory for more information
+ *
+ * SPDX-License-Identifier: MIT
+ */
+
+#include <onyx/process.h>
+#include <onyx/vm.h>
+
+// TODO: Export this stuff in some header, and avoid sticking everything into vm.cpp
+vm_area_struct *vm_search(struct mm_address_space *mm, void *addr, size_t length)
+    REQUIRES_SHARED(mm->vm_lock);
+
+static long do_pagemap(struct mm_address_space *as, unsigned long start, unsigned long end,
+                       u64 *pagemap)
+{
+
+    scoped_mutex g{as->vm_lock};
+    long pfns_processed = 0;
+    struct vm_area_struct *vma = vm_search(as, (void *) start, 1);
+    if (!vma)
+        return -ENOMEM;
+
+    while (vma)
+    {
+        if (vma->vm_start > end)
+            break;
+
+        spin_lock(&as->page_table_lock);
+
+        /* Note: This is bad and slow(er). Nonetheless, it does the job for now */
+        for (; start < end; start += PAGE_SIZE)
+            pagemap[pfns_processed++] = __get_mapping_info((void *) start, as);
+
+        spin_unlock(&as->page_table_lock);
+
+        if (start == end)
+            break;
+
+        vma = containerof_null_safe(bst_next(&as->region_tree, &vma->vm_tree_node),
+                                    struct vm_area_struct, vm_tree_node);
+    }
+
+    return pfns_processed;
+}
+
+int sys_mpagemap(void *addr, size_t length, u64 *pagemap)
+{
+    unsigned long start = (unsigned long) addr & -PAGE_SIZE;
+    length += (unsigned long) addr & (PAGE_SIZE - 1);
+    unsigned long end = ALIGN_TO(start + length, PAGE_SIZE);
+    struct mm_address_space *as = get_current_address_space();
+    struct page *buffer;
+    int ret = 0;
+
+    if (start < as->start || end > as->end)
+        return -EINVAL;
+
+    buffer = alloc_page(GFP_KERNEL);
+    if (!buffer)
+        return -ENOMEM;
+
+    while (start < end)
+    {
+        /* Limit runs to PAGE_SIZE / sizeof(u64) */
+        u64 *kbuffer = (u64 *) PAGE_TO_VIRT(buffer);
+        unsigned long this_run_end = cul::min(end, start + (PAGE_SIZE / sizeof(u64)));
+        long processed = do_pagemap(as, start, this_run_end, kbuffer);
+
+        if (processed == 0)
+            break;
+
+        if (processed < 0)
+        {
+            ret = processed;
+            break;
+        }
+
+        if (copy_to_user(pagemap, kbuffer, processed * sizeof(u64)) < 0)
+        {
+            ret = -EFAULT;
+            break;
+        }
+
+        pagemap += processed;
+        start += processed << PAGE_SHIFT;
+    }
+
+    free_page(buffer);
+
+    return ret;
+}